Hi.
I want to update sphinx to let it be able to recognize Urdu language. Does Sphinx have an acoustic model for Urdu language?
If not, my understanding is that:
I have two options:
1. Create a training corpus. But creating a phonetic dictionary for Urdu is the most difficult task. Does Sphinx have an Urdu phonetic dictionary? Or if anyone has used one, can he tell me where to find a Roman-Urdu dictionary with pronunciation (phonetic dictionary for Urdu, written in English alphabets)?
2. I can buy an acoustic model and phonetic dictionary for Urdu from somewhere. Do you know of any resource from where I can purchase this, or get it created on order?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
السلام علیکم و رحمۃاللہ و برکاتہ
عزیزی طیب عتیق صاحب
گزارش ہے کہ اپنے تجربات سے اس ناچیز کو بھی آگاہ فرمائیں۔
آپ کا تجربہ کہاں تک پہنچا؟
اس سلسلے میں ، میں کیا مدد کرسکتا ہوں؟
کیا کوئی نان پروگرامیبل جی یو آئی مل سکتا ہے ،،،اس کام کے لئے،،،
یعنی کچھ ایسا کہ ہم اس میں ڈیٹا جمع کر سکیں اور اسےآزما بھی سکیں۔
جیسا کہ ڈریگن اسپیچ ریکگنیشن پروفیشنل میں ہوتا ہے۔
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
But creating a phonetic dictionary for Urdu is the most difficult task. Does Sphinx have an Urdu phonetic dictionary? Or if anyone has used one, can he tell me where to find a Roman-Urdu dictionary with
I don't think it's a huge problem, you just map letters to sounds with simple rules and usually it's enough. You can implement rules in your favorite scripting language like Python. You can find more details in extensive research on the subject, for example the following thesis:
I can buy an acoustic model and phonetic dictionary for Urdu from somewhere. Do you know of any resource from where I can purchase this, or get it created on order?
Yes, you can purchase the corpus, for example here:
Or you can collect the corpus yourself, it will be cheaper but will require more work. You can use TV Shows, movies with subtitles and audiobooks to collect extensive database.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
So I'm assuming that Urdu letters are not supported by the phonetic dictionary.
you just map letters to sounds with simple rules
But I thought that sounds of different words vary in different words? For example, "O" in these pronunciations from default.dic:
'quote K W OW T
world W ER L D
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The sampling rate here is 44.1K. Does resampling a sample work with acoustic model adaptation?
For Urdu it's easier as you an see in thesis I shared above.
(Whoa, that's my university, and the Co-Supervisor is my teacher)
You're right. It might be easier. But still there are words in Urdu too. Like:
بالکل
Pronunciation by the rules: BALKL
Correct pronunciation: BILKUL
This comes from the rule "While at nucleus positions they make long vowels". I guess I'll have to do something about that.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi.
I want to update sphinx to let it be able to recognize Urdu language. Does Sphinx have an acoustic model for Urdu language?
If not, my understanding is that:
I have two options:
1. Create a training corpus. But creating a phonetic dictionary for Urdu is the most difficult task. Does Sphinx have an Urdu phonetic dictionary? Or if anyone has used one, can he tell me where to find a Roman-Urdu dictionary with pronunciation (phonetic dictionary for Urdu, written in English alphabets)?
2. I can buy an acoustic model and phonetic dictionary for Urdu from somewhere. Do you know of any resource from where I can purchase this, or get it created on order?
السلام علیکم و رحمۃاللہ و برکاتہ
عزیزی طیب عتیق صاحب
گزارش ہے کہ اپنے تجربات سے اس ناچیز کو بھی آگاہ فرمائیں۔
آپ کا تجربہ کہاں تک پہنچا؟
اس سلسلے میں ، میں کیا مدد کرسکتا ہوں؟
کیا کوئی نان پروگرامیبل جی یو آئی مل سکتا ہے ،،،اس کام کے لئے،،،
یعنی کچھ ایسا کہ ہم اس میں ڈیٹا جمع کر سکیں اور اسےآزما بھی سکیں۔
جیسا کہ ڈریگن اسپیچ ریکگنیشن پروفیشنل میں ہوتا ہے۔
I don't think it's a huge problem, you just map letters to sounds with simple rules and usually it's enough. You can implement rules in your favorite scripting language like Python. You can find more details in extensive research on the subject, for example the following thesis:
https://www.cs.cmu.edu/~araza/Automatic_Speech_Recognition_System_for_Urdu.PDF
Yes, you can purchase the corpus, for example here:
http://www.speechocean.com/en-ASR-Corpora/2038.html
Or you can collect the corpus yourself, it will be cheaper but will require more work. You can use TV Shows, movies with subtitles and audiobooks to collect extensive database.
So I'm assuming that Urdu letters are not supported by the phonetic dictionary.
But I thought that sounds of different words vary in different words? For example, "O" in these pronunciations from default.dic:
'quote K W OW T
world W ER L D
In English rules are more complex than in most languages of the world. For Urdu it's easier as you an see in thesis I shared above.
The sampling rate here is 44.1K. Does resampling a sample work with acoustic model adaptation?
(Whoa, that's my university, and the Co-Supervisor is my teacher)
You're right. It might be easier. But still there are words in Urdu too. Like:
بالکل
Pronunciation by the rules: BALKL
Correct pronunciation: BILKUL
This comes from the rule "While at nucleus positions they make long vowels". I guess I'll have to do something about that.
You can always resample
You can also crawl pronunciations from Wiktionary
http://en.wiktionary.org/wiki/بالکل#Urdu
They have quite many words transcribed there. After that you can train g2p classifier to predict pronunciation for the rest.
All those possibilities are described in dictionary tutorial
http://cmusphinx.sourceforge.net/wiki/tutorialdict
The crawl task for Wiktionary stands for quite some time.
i can help you
conact me on hiyassat at gmail dot com